AGENT SKILL

embeddings-retrieval

Use UniLM E5 and SimLM for embedding evaluation, dense retrieval, reranking, hard-negative mining, teacher scoring, and retrieval training workflows.

License 许可
MIT
最近更新
2026-09-03
Stars
★ 351
Forks
27

01 前言 Front Matter

name
embeddings-retrieval
license
MIT
metadata
{ "disco-role": "operating" }
description
Use UniLM E5 and SimLM for embedding evaluation, dense retrieval, reranking, hard-negative mining, teacher scoring, and retrieval training workflows.
disable-model-invocation
true

02 资源文件 4

03 SKILL.md 正文

Embeddings Retrieval

Use this sub-skill when a task names E5, SimLM, BEIR, MTEB, MS MARCO, DPR/NQ retrieval, dense corpus/query encoding, reranking, hard-negative mining, or teacher-score generation.

For general sequence-to-sequence generation, route to language-seq2seq. For image-text or document AI retrieval that does not explicitly name E5 or SimLM, route to vision-document-ai. For PFPO/ReSA math evaluation or generic training architecture questions, route to architectures-training.

Fast Routing

  • Choose E5 for off-the-shelf text embeddings, BEIR retrieval evaluation, non-retrieval MTEB tasks, multilingual embeddings, and prefix/pooling questions.
  • Choose SimLM biencoder for dense passage retrieval pipelines that separately encode a corpus, encode/search queries, and optionally mine hard negatives from search output.
  • Choose SimLM reranker for cross-encoder scoring of (query, candidate passage) pairs from an existing top-k run; do not use it to encode a corpus.
  • Choose SimLM RLM/KD training for replacing-language-model pretraining, distillation data generation, or retraining MS MARCO/DPR retrieval models.

Bundled References and Script

  • Read references/model-and-data-reference.md to choose E5/SimLM model families, prefixes, pooling, dependencies, file roles, and data layouts.
  • Read references/workflows.md to construct safe E5 evaluation commands and plan SimLM encode/search/rerank/train/teacher-score phases.
  • Read references/troubleshooting.md when prefixes, multilingual flags, GPU memory, downloads, DeepSpeed, or missing SimLM data files break a workflow.
  • Run scripts/build_e5_eval_command.py --help to print a validated E5 BEIR or MTEB command without downloading data, loading a model, or launching evaluation.

Safety Defaults

  • Treat E5 BEIR/MTEB evaluation as network- and GPU-heavy: the upstream scripts download benchmark data/model weights and can run for hours.
  • Treat SimLM download, train, encode, search, rerank, and teacher-score scripts as reference-only unless the user has explicitly provided data, GPUs, storage, and permission for long jobs.
  • Prefer command construction, file validation, and dry-run planning before any native run. SimLM code asserts CUDA availability for core training/inference paths.
  • Keep generated commands rooted in a caller-provided working copy or environment; this skill does not require or expose the original source checkout.

04 同仓库更多 Skills

VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-triton-opt

Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-solve

END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-skill

PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-pattern-fusion

Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-orchestrate

Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-feedback

Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).

★ 351